lora-training / kokona /README.md
khanon's picture
initial commit
7529c6f
|
raw
history blame
1.82 kB
# Sunohara Kokona (Blue Archive)
## Usage
Use any or all of these tags to summon Kokona:
`1girl, halo, animal ears, orange eyes, long hair, grey hair`
She really doesn't need many, just `1girl, halo` is enough to summon an instantly recognizable version of her. The model may be a little overtrained but it looks good.
For her normal outfit:
`china dress, vertical-striped dress, white skirt, frilled skirt, off-shoulder, jacket`
Somewhat frequently, a strange, mysterious man in a black suit likes to appear out-of-frame beside her. You can shoo him away with `1boy` in the negative prompt. Not sure why this happens as there was nothing like it in the training data.
The AI is generally pretty good at picturing her, but it really struggles with her pout. Try various combinations of `pout`, `:t`, `pouting` and negative prompt `wavy mouth` to try to get it to appear correctly, or just use img2img without the LoRA. It seems to be related to the LoRA since vanilla NAI can do pouting just fine.
Her stamp is in the training data and it sorta knows what it is but is probably too small to be recognized by the AI.
You can add or ignore `kokona, blue archive`; while they were in her captions, their effects aren't especially strong.
Weights close to 1 work well with her, she's very cooperative.
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 95 curated images, repeated 7 times (665 total images / 3 batchsize = ~222 iterations per epoch, 3 epochs)
- Dataset included a mixture of SFW and NSFW.
- 1.5e-5 learning rate, same as Iroha.
- Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
- `kokona, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- No variants, she's well-behaved.